Roster mapping each role (researcher, curator, reviewer, copy editor,
picker, conductor) to its pipeline stage and canonical prompt file,
the shared house rules, the day-in-the-life commit flow, and the
copy editor's subagent brief verbatim — the one prompt not stored as
a skill; daily-pipeline stage 4 now references it as canonical.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
One written standard for every audience-facing word: concise, to the
point, inspiring delight through specificity — with word budgets per
copy type (reviews had drifted from a ~370-word median to 500-1000+).
All writing skills now reference the charter, and daily-pipeline gains
a stage-4 copy-edit pass: a subagent that didn't write the day's copy
tightens it against the charter, prose only, diff reviewed before
commit. Digest gains a copy-edit line.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
The real failure mode is 0-1 images (wall of text); two heroes read
fine on the 1-2 column layouts. Keep ~4 as the aim — prefer
hero-bearing candidates among equal fits, more images for bigger
bundles — since 2 among 8 items still looks thin on the wide
4-column grid. Restores the 3-item minimum bundle size.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
Bundles now ship only when at least four of their items have a hero.*
image, purely so the grid never renders image-thin. Clusters that miss
the floor are held back and logged rather than padded or written thin;
the weekly catch widens its window to older unbundled hero-bearing
finds instead of holding. Capture gets a nudge to try a second image
source before giving up on a hero.
Claude-Session: https://claude.ai/code/session_01WZaczDJjL3xZ3u5spsN5AL
- daily-finds: feed-first fetch strategy (fetch: feed + feedUrl), cadence-aware
scanning, new score-and-shortlist stage (shortlist: true, up to 12/day, top
1-3 flagged as review candidates), Amazon affiliate lookup for shortlisted
physical products, and authority to apply mechanical sources.json maintenance
(skip dead sources, demote dormant ones, record working fetch paths) instead
of logging recommendations nobody executes.
- build-bundle: auto mode (no confirmation gate when unattended) and a 'weekly'
date-based catch mode so shortlisted finds never sit invisible for weeks.
- cluster-bundles: auto mode.
- build-review: hero.ts registration step removed (heroes auto-discovered from
hero.<ext>); mandatory fact-check against the product's own site in auto
mode; amazonLink carry-through; auto mode publishes and logs instead of
pausing twice for approval.
- New daily-pipeline skill: capture → bundle → reviews → pick of the day →
build → commit → push → digest, with granular per-stage commits. This is the
entrypoint for scheduled runs.
- Project CLAUDE.md documenting the pipeline and its foot-gun conventions.
rzen